17/07/2024 Erik Kusch
11
Erik Kusch
Senior Engineer
Machine Readable Nature Research Group (MANA)
Department of Research and Collections
Natural History Museum
University of Oslo
17/07/2024 Erik Kusch
STRINGS BETTER LEFT TO
FRAY? - ECOLOGICAL
NETWORK INFERENCE
Investigating Inference Accuracy using Demographic Simulations
ISEC 2024
17/07/2024 Erik Kusch
2Methods of Ecological Network Inference
ISEC 2024
COOCCUR NETASSOC HMSC NDD-RIM
Multiple frameworks for association/interaction inference have been proposed
COOCCUR
NETASSOC
HMSC
NDD-RIM
Distinction of network inference approaches:
Consideration of Environmental Conditions:
Environmental conditions affect expressions of interactions in identity
and magnitude
Spectrum of Co-occurrence–Performance:
Information content changes when considering presence/absence,
abundance, or performance
Consideration of
Environmental Conditions
Explicit
Implicit
None
Biodiversity Input Data Type
Co-Occurrence Abundance Performance
COOCCUR
NETASSOC
NDD-RIM
HMSCHMSCHMSC
HMSCHMSC
HMSC
NETASSOC
17/07/2024 Erik Kusch
3Inference Dissimilarity & Drivers
Consideration of environmental conditions
a driving factor of inference outcome at
macro scale?
Performance information may stabilise
inference outcome.
But how do we know which inference approach
yields the most accurate results?
ISEC 2024
17/07/2024 Erik Kusch
44
INVESTIGATING INFERENCE
ACCURACY USING DEMOGRAPHIC
SIMULATIONS
4
Research Question
How accurate is network inference?
ISEC 2024
17/07/2024 Erik Kusch
5Study Concept & Simulation Set-Up
Need to know the “true” ecological network to evaluate network
inference accuracy
Random generation of ecological network
Biodiversity data analysed by the assessed inference approach
ought to reflect:
Environmental conditions & species-specific niche preferences
Interactions between individuals of interacting species
Demographic simulation with variable death rate:
ISEC 2024
17/07/2024 Erik Kusch
6Variable Death Rate Components
ISEC 2024
17/07/2024 Erik Kusch
7Network Realisation
True Potential Network True Realised Network
Environmental Preference Similarity
ISEC 2024
17/07/2024 Erik Kusch
8Network Inference
True Realised Network
Occurrence (O) –50% Abundance (A) –50% Performance (P) –30%
[O]Clim –30% [A]Clim –50% [P]Clim –30%
HMSC-Inferred Networks
ISEC 2024
17/07/2024 Erik Kusch
9Inference Accuracy
1. Network inference accuracy varies.
2. Gains to network inference accuracy when
considering abundance data and environmental
conditions.
Non-Abiotic Parameterisation Abiotic Parameterisation
All Parameterisations
Missed Absent Rate
Missed Negative Rate
Missed Positive Rate
True Absent Rate
True Negative Rate
True Positive Rate
ISEC 2024
How accurate is network inference?
What
Now?
Determine real-world
parameterisation for the simulation
framework.
17/07/2024 Erik Kusch
10 Inference Likelihood II What drives differences in
network inference accuracy?
[O] [A]Clim
Positive Associations
Negative Associations
[A]
3. Inference (irrespective of whether it is correct or not) of a positive association strongly depends on its strength and the
differences in environmental preference of the association partners. Negative associations are rarely inferred.
ISEC 2024
17/07/2024 Erik Kusch
11 Inference Accuracy II What drives differences in
network inference accuracy?
[O] [A]Clim
Positive Associations
Negative Associations
[A]
4. Correct inference of an association (irrespective of sign) strongly depends on its strength and the differences in
environmental preference of the association partners.
ISEC 2024
17/07/2024 Erik Kusch
12 Study Design & Data Streams
# 810,919 at 160401 plots
# 70,797 at 640 plots
Yosemite Forest Dynamics Plot
Yosemite National Park Temperate Conifer
Forest Biome
Pre-fire event in 2013
Raw
Data
Subsetting in
Time and Space
# 34,444 at 640 plots
11 species
Plot-Scale
# 291 at 101 plots
13 species
Region-Scale
# 96,169 at 46,328 plots
15 species
Macro-Scale
BIEN
YFDP Traits
KrigR
V.PhyloMaker
Plot Temperature Soil
Moisture
Precipi-
tation
Evapo-
ration
… … … … …
… … … … …
Species SLA Leaf Carbon Leaf Nitrogen
… … … …
… … … …
Plot Species 1 Species 2 Species …
… … … …
… … … …
Phylogeny
Distributional Null
Expectation
Environmental
Conditions
Spatial Products
Functional Trait
Data
Species Plot Performance Abundance Presence/ Absence
… … … … …
… … … … …
Adding Data for Use in
Ecological Network Inference
ISEC 2024
17/07/2024 Erik Kusch
13 No Consensus of Inference Across Approaches
I
Does choice of inference method
affect
inferred network structure?
ISEC 2024
17/07/2024 Erik Kusch
14 No Consensus Across Scales
II
Are inference methods and
their networks scalable?
NETASSOC
HMSC
COOCCUR
NDD-RIM
ISEC 2024
